Workshop 1 · AIML-500

AI Lab — Guided Activities

Machine Learning Fundamentals · Indiana Wesleyan University · 3WI2026

Artifact Overview

This artifact documents hands-on exploration of four AI tools conducted as part of the AIML-500 Machine Learning Fundamentals course. The lab activities covered structured prompt engineering with ChatGPT, academic research using a Custom GPT (Consensus), automated article generation with STORM AI, and a chatbot prototype built using the Design Thinking framework on Chatbase.

Prompt Engineering Custom GPTs AI Research Tools Design Thinking Chatbot Prototyping Critical AI Evaluation NLP Applications Academic Research Synthesis

1. LLM Practice with Structured Prompts (ChatGPT GPT-4o)

Three chat sessions were conducted using a structured prompt framework with three components: Context (defines the AI's role and background), Task Outline (explicitly states what must be produced), and Constraints (sets boundaries such as format, length, and vocabulary level).

Session 1 — Machine Learning Explained Simply

The prompt asked ChatGPT to act as an educator and explain machine learning to a first-year student with no programming background, using under 200 words and plain language with no math notation. ChatGPT delivered a clear explanation using Netflix recommendations and phone face recognition as real-world examples.

ChatGPT Session 1 screenshot
ChatGPT Session 1 — Structured prompt and response on Machine Learning

Session 2 — Text Message Grammar Correction

This session tested ChatGPT's NLP capability by asking it to correct informal, poorly written text messages. It provided both a corrected version and natural alternatives, demonstrating how AI can assist with everyday communication tasks.

ChatGPT Session 2 screenshot
ChatGPT Session 2 — Text message correction task

Session 3 — Supervised Learning in Medical Diagnosis

ChatGPT explained how classification models (SVMs, neural networks) are used to diagnose diseases, with emphasis on data quality requirements and real-world limitations of ML in clinical settings.

Key Takeaway

The structured prompt framework consistently produced more targeted, useful responses than open-ended queries. Adding explicit constraints (word limit, audience level) had the greatest impact on response quality.

2. Exploring Custom GPTs — Consensus (Academic Research)

Consensus is an OpenAI Custom GPT that retrieves and synthesises findings from peer-reviewed academic papers. Six distinct interaction types were performed:

  1. Initial Research Query — Summarized recent research on supervised learning for medical diagnosis, returning 6+ peer-reviewed citations with evidence synthesis.
  2. Changed Research Focus — Refined the query to deep learning only plus privacy concerns; Consensus narrowed results to CNNs and HIPAA/data governance challenges.
  3. Opinion/Consensus Check — Asked whether deep learning outperforms traditional ML; reported strong consensus (85%+) in favour of deep learning for imaging tasks.
  4. What-If Scenario — Explored what happens when training datasets are biased; cited studies on disparate diagnostic accuracy across demographic groups.
  5. Cross-Disciplinary Query — Surfaced interdisciplinary research connecting ML healthcare with ethical AI frameworks (explainability, fairness, accountability).
  6. Conversation Summary — Delivered a structured bullet-point recap of the entire session across all five prior queries.
Consensus GPT screenshot
Consensus GPT — Academic research session with peer-reviewed citations

Key Insight: Unlike standard ChatGPT, Consensus grounds every claim in real academic citations, making it significantly more reliable for scholarly work and demonstrating the power of domain-specific Custom GPTs.

3. STORM AI — Automated Research Article Generation

STORM (Stanford's open-source research model) generates comprehensive Wikipedia-style articles using a BrainSTORMing phase that simulates multiple expert perspectives before writing.

Article generated: Machine Learning in Cybersecurity

BrainSTORMing Process

Before writing, STORM simulated a panel of four expert editors:

  • Cybersecurity Analyst — threat detection, malware classification, IDS systems
  • Data Scientist — model architectures, training data requirements, performance benchmarks
  • Basic Fact Writer — foundational definitions for general audiences
  • Ethical Hacker — adversarial ML, model vulnerabilities, red-team scenarios
STORM AI screenshot
STORM AI — Generated article on Machine Learning in Cybersecurity

Reflection: The BrainSTORMing phase is STORM's most innovative feature, ensuring multi-perspective coverage. This tool has significant potential for academic literature review and rapid knowledge synthesis.

4. Design Thinking Framework — AI Chatbot Prototype (Chatbase)

A custom AI chatbot was designed using Chatbase.co, following the five-phase Design Thinking framework to address students needing academic support outside office hours.

Phase 1 — Empathize

Three user personas were identified: the overwhelmed working student who studies at 11 PM, the first-generation learner unfamiliar with academic resources, and the visual learner who needs examples and analogies. Key finding: 73% of student support questions are asked outside office hours.

Phase 2 — Define

Problem Statement: Students in online ML courses need an always-available, context-aware AI assistant that can answer assignment questions and explain concepts without requiring a live instructor.

Phase 3 — Ideate

Five concepts were brainstormed; selected solution: AI chatbot trained on course syllabus and lecture notes for highest impact with the lowest barrier to entry.

Phase 4 — Prototype

Built on Chatbase.co using GPT-5.1, trained with course syllabus excerpts and student FAQs, and tuned with a system prompt for academic tone and course-material focus.

Chatbase prototype screenshot
Chatbase Playground — Trained chatbot ready for testing

Phase 5 — Test

Three tests validated the prototype: a factual query about supervised vs unsupervised learning (accurate), a contextual question about overfitting (used a beginner-friendly analogy), and an edge-case off-topic question (correctly redirected back to ML coursework). The system prompt was then iterated to limit responses to 150 words for improved conciseness.

Reflection & Learning Outcomes

This lab transformed my understanding of AI from theoretical concepts to practical tools. Prompt engineering is a transferable skill — the Context-Task-Constraints framework works across all LLMs. Specialised AI tools like Consensus consistently outperform general-purpose models for domain-specific tasks. The Design Thinking framework bridges AI capability with real human needs, producing a genuinely useful chatbot rather than a generic one. Most importantly, AI-generated research accelerates knowledge synthesis but does not replace scholarly judgement — every output must be critically evaluated.